3D simultaneous localization and mapping in texture-less and structure-less environments using rank order statistics

The area of mobile robotics has attracted substantial attention from researchers around the globe. This has resulted in many technological advances and breakthroughs in this area. Nowadays, mobile robots have become widely available, such as robotic vacuum cleaners, those working in industrial settings, hospitals, warehouses, office environments, and military and space applications. Many of the tasks that robots are assigned to complete require them to be fully autonomous, i.e. without any human intervention. For this reason, it is of extreme importance that a robot is able to localize itself and map the environment accurately. These tasks are particularly necessary, if the robot is required to navigate in an unknown dynamic environment, plan its path and avoid collisions with other objects. Simultaneous Localization and Mapping (SLAM) is the process in which a robot is required to simultaneously localize itself in an unknown environment and incrementally build a map of its surroundings. The SLAM problem has been studied extensively and many solutions have been proposed to solve this problem. Most recently, due to availability of low cost RGB-D cameras, which provide RGB images as well as pixel depth information, 3D SLAM has received a large amount of interest and is used in various applications. The additional information provided by 3D maps can be used to improve navigation, collision avoidance and path planning algorithms. The research presented in this thesis, intends to study the 3D SLAM problem in difficult situations, such as a robot navigating in areas in which the visual sensing of the robot is hindered. For example, firefighting robots and those performing search and rescue operations, may need to perform SLAM in smoky and dark environments. We propose a method that uses an RGB-D sensor to register images in dark scenes. This method utilizes both the RGB and IR images provided by an RGB-D sensor. Another challenging scenario, is the need to perform SLAM in an environment containing limited texture and structure information. These environments are commonly encountered in offices, warehouses and residential buildings. Many of the aforementioned environments contain texture-less and structure-less walls, floors and ceilings. This poses a significant challenge that is to match the frames, estimate the motion and map the environment. We present methods that utilize both geometric and texture information to overcome the aforementioned challenges. At the core of the proposed methods, we develop Rank Order Statistics based on informative sampling techniques that are able to sample the dense depth points into a subset of points, carrying highly useful information for registration. In addition, we outline one of the main limitations of using RGB-D cameras for SLAM applications: the limited field of view (FOV). In contrast, monocular SLAM systems can exploit wide-angle cameras and do not have the depth range limitation. As such, we present a SLAM method that fuses the information obtained from both an RGB-D camera and a wide-angle monocular camera. We show that this is particularly beneficial for large-scale 3D reconstruction of indoor environments. Finally, we propose a image based feature exaction method that is able to find a large number of highly repeatable features. We show that when pairing these features with a robust image descriptor such as SIFT, they become highly invariant to various image transformations. We also show that they could be highly useful for visual odometry applications, as well as for obtaining denser looking 3D models when performing monocular SLAM/Structure from Motion